Enhancing YOLOF with attention mechanism for improved object detection performance

Abstract

Object detection is one of the central problems in computer vision, particularly in fields such as robotics and autonomous driving, where accurate perception and understanding of the surrounding environment is a prerequisite for safe and efficient operation. In traditional approaches, object detection models often require significant computational costs and lengthy training times, which hinder practical deployment. To address these challenges, an improved version of the YOLOF (You Only Look One-level Feature) model is proposed, focusing on enhancing efficiency and accuracy in object detection. The proposed method strengthens the backbone with an attention mechanism to refine feature representations, enabling the model to better focus on important regions in the image. This approach preserves the inherent simplicity and speed of YOLOF while improving detection performance. The model is evaluated on two benchmark datasets: MS COCO and BDD100K. On MS COCO, the model achieves an AP of 40.4%, ranking among the top-performing one-stage detection methods. On BDD100K, a dataset designed for autonomous driving scenarios, the model achieves an AP of 33.2%, demonstrating strong adaptability in complex and diverse environments. These results confirm that the proposed method strikes a balance between accuracy and efficiency, making it a suitable choice for real-world applications where computational resources are limited.